Paper. Series. Bernard Salanié. Discussion

Transcription

1 Columbia University Department of Economics Discussion Paper Series Matching with Trade-offs: Revealed Preferences over Competing Characteristics Alfred Galichon Bernard Salanié Discussion Paper No.: Department of Economics Columbia University New York, NY March 2010

2 Matching with Trade-offs: Revealed Preferences over Competing Characteristics Alfred Galichon 1 Bernard Salanié 2 First version dated December 6, The present version is of March 22, Economics Department, École polytechnique; alfred.galichon@polytechnique.edu 2 Department of Economics, Columbia University; bsalanie@columbia.edu. 3 The authors are grateful to Guillaume Carlier, Pierre-André Chiappori, Piet Gauthier, Jerry Hausman, Jim Heckman, Kevin Lang, Guy Laroque, Christian Léonard, Sonia Oreffice, Rob Shimer as well as seminar participants at Crest, Ecole Polytechnique, séminaire Roy, University of Chicago, Harvard-MIT, the Toulouse School of Economics, the University of New South Wales, the University of Melbourne, Australian National University and the University of Alicante for useful comments and discussions. Much of this paper was written while Salanié was visiting the Toulouse School of Economics; he is grateful to the Georges Meyer endowment for its support. Galichon gratefully acknowledges support from Chaire EDF-Calyon Finance and Développement Durable, and Chaire Axa Assurance et Risques Majeurs and FiME, Laboratoire de Finance des Marchés de l Energie (

3 Abstract We investigate in this paper the theory and econometrics of optimal matchings with competing criteria. The surplus from a marriage match, for instance, may depend both on the incomes and on the educations of the partners, as well as on characteristics that the analyst does not observe. The social optimum must therefore trade off matching on incomes and matching on educations. Given a flexible specification of the surplus function, we characterize under mild assumptions the properties of the set of feasible matchings and of the socially optimal matching. Then we show how data on the covariation of the types of the partners in observed matches can be used to estimate the parameters that define social preferences over matches. We provide both nonparametric and parametric procedures that are very easy to use in applications. Keywords: matching, marriage, assignment. JEL codes: C78, D61, C13.

4 Introduction Starting with Becker (1973), most of the economic theory of one-to-one matching has focused on the case when the surplus created by a match is a function of just two numbers: the one-dimensional types of the two partners. As is well-known, if the types of the partners are one-dimensional and are complementary in producing surplus then the socially optimal matches exhibit positive assortative matching. Moreover, the resulting configuration is stable, it is in the core of the corresponding matching game, and it can be implemented by the celebrated Gale and Shapley (1962) deferred acceptance algorithm. While this result is both simple and powerful, its implications are also quite unrealistic. If we focus on marriage and type is education for instance, then positive assortative matching has the most educated woman marrying the most educated man, then the second most educated woman marrying marrying the second most educated man, and so on. In practice the most educated woman would weigh several criteria in deciding upon a match; even in the frictionless world studied by theory, the social surplus her match creates may be higher if she marries a man with less education but, say, a similar income. Income and education are only imperfectly correlated; and the correlation patterns differ for men and women. Then the optimal match must trade off assortative matching along these two dimensions. This point is quite general: with multiple types, the stark predictions of the one-dimensional case break down. Analysts of matching have long felt the need to accommodate the imperfect assortative matching observed in the data, of course. One possibility is to introduce search frictions, as in Shimer and Smith (2000); but the resulting model is hard to handle, and it still implies assortative matching, under stronger conditions. In our view, a simpler solution consists in allowing the joint surplus of a match to incorporate heterogeneity that is unobserved by the analyst. As explained below, this was pioneered by Choo and Siow (2006) and extended by Chiappori, Salanié, Tillman, and Weiss (2008); a different variant, used by Chiappori, Oreffice, and Quintana-Domeque (2009), assumes that the observed heterogeneity has a one-dimensional index structure. Our contribution here is twofold. First, we exhibit a very 1

5 simple nonparametric estimator of the surplus function that is at the heart of matching models with transferable utility; second, we explore the properties of optimal or equilibrium matchings 1 and we use our results to describe an empirical strategy and to obtain parametric estimators. For simplicity, we use the language of the economic theory of marriage in our illustrations; yet nothing we do actually depends on it. The methods proposed in this paper apply just as well to any one-to-one matching problem or bipartite matchings. In fact, we can even extend them to problems in which the sets of partners are determined endogenously as with same-sex unions. extensions of our setting. This is investigated in Section 8, where we consider possible We do require, however, that utility be transferable across partners. function is indeed the surplus created by a match, defined as Our primitive Φ( x, ỹ) where x is the full type of a man and ỹ the full type of the woman who is his partner in this match. A full type has components that are observed by the econometrician, and others that are only observed by the participants on the market. We denote x the observable type of a man, and y that of a woman. As is well-known, this model is too general to be empirically testable: even without unobserved heterogeneity (when x coincides with x and y coincides with ỹ), any observed assignment can be rationalized by a well-chosen surplus function. This is a consequence of a more general theorem by Blair (1984). Echenique (2008) shows that on the other hand, some collections of matchings are not rationalizable: if the analyst can observe identical populations on several assignments, then these assignments must be consistent with each other in a sense that his paper makes precise. But we are unlikely to have such data at hand in general. 1 A word on terminology: like most of the literature, we call a match the pairing of two partners, and a matching the list of all realized matches. 2

6 Relatedly, analysts sometimes observe several subpopulations which are matching independently and yet have the same surplus function. Fox (2009) shows that under a rankorder condition on the unobserved heterogeneity, it is then possible to identify several important features of the surplus function, and in particular how important complementarity is on various dimensions. While analyzing complementarity is also one of our goals here, many of the applications we have in mind do not fit Fox s assumption that there be enough variation across subpopulations with identical surplus. Marriage markets, for instance, seem to be either so disconnected that their surplus functions are unlikely to be similar, or too connected to make it possible to ignore matching across markets. In this paper, we only assume that we have data on one instance of a matching problem, such as the marriage market in the US in the 1980s, or the market for CEOs. Our data will consist of the values of the observable types of both partners in each realized match, and of the types of unmatched individuals. Since the optimal/equilibrium matching is determined on the basis of both the observable and the (to us) unobservable types, we will need to impose assumptions that allow us to integrate over the distribution of the unobservable types in a manageable way. Our aim is to start from the observable matching (the distribution of matches across observable types) and to recover as much information as we can on the surplus function. That is, the econometrician observes a distribution π(x, y) over the observable types of both partners in observed matches; and he seeks to recover Φ. To do this, we have to impose assumptions. First, we restrict our analysis to the case when observable types take a finite number of values. Then we impose a separability assumption that excludes interactions between the unobservable types of the partners in the production of joint surplus: Φ( x, ỹ) = Φ(x, y) + χ( x, y) + ξ(ỹ, x), where the χ s and ξ s have conditional mean zero. One interpretation is that the surplus created by a potential match is an unknown function of the types of the partners only, plus preference shocks that are observed by all participants but not by the analyst in the 3

7 nature of unobserved heterogeneity. This separability assumption was used by Choo and Siow (2006) and then generalized by Chiappori, Salanié, Tillman, and Weiss (2008), who showed that the matching equilibrium then boils down to a series of parallel discrete choice models. While this is an important step on the way to a solution, the resulting model is still too rich to be taken to the data. We need to restrict the distribution of unobserved heterogeneity further. To do this, we adopt Choo and Siow (2006) s assumption that the terms χ and ξ above are type-i extreme values giving the model the structure of a multinomial logit. If the analyst is lucky enough to have very rich data, unobserved heterogeneity is almost irrelevant and the observable matching π maximizes the observable surplus function. A bit more formally, let P and Q denote the marginal distributions over observed types of men and women respectively. Then if there is no unobserved heterogeneity, the observed matching π must maximize E π Φ(X, Y ) over the set M(P, Q) of all joint distributions π that have marginals P and Q. On the other hand, if data is so poor that unobserved heterogeneity dominates, then the analyst should observe something that, to him, looks like completely random matching: π should be the product P Q of the marginal distributions. As is well known, independence maximizes relative entropy: π = P Q minimizes the mutual information over M(P, Q). I(π) = E π log π(x, Y ) P (X)Q(Y ) We show that under our assumptions, for any intermediate amount of unobserved heterogeneity, the observable matching π maximizes a straightforward linear combination of the observable surplus and of the mutual information above: E π Φ(X, Y ) σe π log π(x, Y ) P (X)Q(Y ), where σ measures the size of unobserved heterogeneity. 4

8 Apart from its simplicity, this objective function has two very nice properties when σ > 0: it is globally strictly concave, and it has an infinite derivative in zero. The former implies that the optimal matching is unique, and the latter that all observable matches have positive probability: 0 < π(x, y) < 1 for all (x, y). Finally, for any given Φ and σ (and (P, Q)) the optimal matching π is much easier to compute than in the homogeneous case: we show that the well-known Iterative Projection Fitting Procedure is easily adapted to the structure of this problem. Since IPFP is a very fast, very stable and very simple algorithm (known to some economists as RAS), we consider this to be another attractive property of our method. Under our assumptions, we prove that the surplus function over observable types Φ is nonparametrically partially identified from the data; and that the features that are identified hold considerable economic interest. In fact, the log-likelihood function of the observed matches is one of the surplus functions that can rationalize the data: Φ log π; and we can use it to test for complementarities between any two observable dimensions of the types of the partners, such as the education of the wife and the income of the husband. We can also identify the relative strengths of such complementarities across different dimensions. While this is a remarkably simple and useful result, even discrete types may take a large number of values, making nonparametric estimation impractical all the more so that π is a joint distribution of types. Parametric analysis will often be necessary in practice. Our IPFP algorithm makes maximum likelihood estimation quite simple even for nonlinear models; but models in which the observable surplus function is linear in the parameters turn out to have quite interesting properties. Consider, for instance, approximating the observable surplus function with a linear expansion over some known basis functions (φ k ), with unknown assorting weights Λ: Φ(x, y) = k Λ k φ k (x, y). If the true model in fact belongs to this class, then all relevant information can be expressed 5

9 in terms of the mutual information I of the joint distribution of types and of the average values of these basis functions φ k across couples, which we call covariations; more formally, the numbers and Î = Eˆπ log ˆπ(X, Y ) ˆp(X)ˆq(Y ) Ĉ k = Eˆπ φ k (X, Y ) are sufficient statistics for estimation of Λ and specification testing. This is a very significant reduction in complexity, from the joint distribution ˆπ(x, y) to just these (K + 1) numbers. We first show that if the true model was generated by the assumed basis functions, then in the mutual information Î should be minimal given the vector of covariations Ĉ, in a sense that we made precise. This gives us a specification test. If the test does not reject the null hypothesis, then there exists a vector of assorting weights Λ for which the optimal matching generates exactly these covariations; and if the true model has positive heterogeneity (σ > 0), this vector is unique up to multiplication by a positive constant. Moreover, we can test that a correctly specified model is homogeneous (σ = 0.) These results lead us to propose a moment matching estimator of the assorting weights that has very desirable properties if the model is correctly specified: it is consistent, asymptotically normal, and asymptotically efficient, and it is also very easy to compute as if maximizes a globally strictly convex function. If the model may have been misspecified, then we can still use this estimator to compute the implied joint distribution of types, and compare it to the nonparametric estimator of π; this gives an additional test for misspecification. Moreover, we can use standard techniques to select among potential sets of basis functions. This paper thus proves both a negative and a positive result. The negative part is that even if we assume separable heterogeneity with a multinomial logit structure, the model still cannot be rejected since we can always rationalize it by the parameters Φ = log π and σ = 1 for instance. The positive part is that given any theory about the way the observable types enter the surplus function (as embodied in a set of basis functions (φ k )), we exhibit well- 6

10 behaved estimators of the unknown parameters Λ; we can test whether heterogeneity σ > 0 is needed to rationalize the data; and we can test whether the basis functions adequately describe matching patterns. Our methods can also be used heuristically, to explore ways to understand what goes on in matching markets and how they change across time and space. Standard statistical techniques could for instance be put to work to find the basis functions that explain the largest share of the variation in the data. Such a methodological stance is reminiscent of revealed preferences in consumer theory; in fact the analogy is very sharp, as the underlying theoretical structure is the same. The theoretical work done by Hatfield and Milgrom (2005) and Chiappori, McCann, and Nesheim (2008) also suggest exploring analogies with auctions and hedonic models, respectively. Our depiction of matching markets of course abstracts from many features of real-world markets. We focus on static, frictionless markets, as in much of the literature on marriage markets. Our data merely consists in the knowledge of who is married to whom at a given date. Most models of matching on job markets, for instance, have adopted a much more dynamic perspective, in which job flows in fact provide a lot of information on the underlying parameters. This is hard to do on many matching markets (it would require, for instance, a good theory of divorce) and we leave this for further work. Another recent trend in the economic literature on matching has been to focus on matching technology, such as platforms; on the marriage market, online dating sites are an example (see e.g. Hitsch, Hortacsu, and Ariely (2010) and Lee (2009).) By construction, features of the matching technology such as frictions are encompassed in our matching surplus function, which thus reflects the social distance between the observable characteristics of two individuals, reflecting technological and social accessibility. In a recent paper, Echenique, Lee, and Shum (2009) take the dual approach to both transferrable utility and non-transferrable utility models of matching, combining homogeneity in preferences and flexible matching frictions. Our view is that at this level of generality, it is hard to choose between the two approaches: any particular application will have to face the standard issue of finding instruments to 7

11 identify frictions and preferences separately (when they do not merge, as may be the case with intercaste marriage for instance.) Section 1 sets up the matching model we study in the paper, along with our assumptions on the specification of the observable surplus and the process that drives unobserved heterogeneity. In section 2 we build on these assumptions to derive our main analytical results, and we prove (partial) nonparametric identification in section 3. Section 4 discusses possible empirical strategies; and it shows that the presence of heterogeneity in fact makes the computation of the optimal matching much easier. Under the assumption that the surplus function Φ is unknown up to a linear parametrization, we give our results a geometric interpretation in section 5. Section 6 introduces our tests and estimators and derives their asymptotic properties; and section 7 illustrates our methods on a subsample of 2008 US Census data. We conclude by sketching extensions of our methods. Since much of what we do uses convexity, we recall some definitions and basic results in Appendix A. All proofs are collected in Appendix B. Finally, we should note that there are close parallels between the analysis we develop in the present paper and familiar notions in thermodynamics and statistical physics. E.g the social utility of a matching evokes (minus) the internal energy of a physical system, and the standard error of unobservable heterogeneity parallels its physical temperature. Since the analogy may prove to be as useful to others as it was to us, we elaborate on it in Appendix C. 1 The Assignment Problem Throughout the paper, we assume that two subpopulations M and W of equal size must be matched, and that utility transfers between partners are unconstrained. Each man (as we will call the members of M) must be matched with one and only one member of W (we will call them women.) Thus we do not model the determination of the unmatched population (the singles) in this paper; we take it as data. We elaborate on this point in our concluding 8

12 remarks. Note also that we assumed bipartite matching: the two subpopulations which define admissible partners are exogenously given. This assumption can also be relaxed; see Section 8. Throughout the paper, we illustrate results on the education/income example sketched in the Introduction, which we denote (ER). 1.1 Population characteristics Each man m has characteristics x m, of which a subset x m is observed by the econometrician. We call x m the full type and x m the observable type. Similarly, each woman w similarly has a full type ỹ w and an observable type y w. We denote P (resp. Q) the distribution of full types x (resp. ỹ) in the subpopulation M (resp. W ), and P (resp. Q) the distribution of observable types x (resp. y.) In observed datasets we will have a finite number N of men and women, so that P and Q are the empirical distributions over the types observed in the sample, {x 1,..., x N } and {y 1,..., y N } respectively. Take the education/income example: there a first dimension of observable types is education E {D, G} (dropout or graduate), and a second dimension is income class R, which takes values 1 to n R. P describes both the number of graduates among men and the distributions of income among graduate men and among dropout men. In addition, full types may incorporate physical characteristics, religion, and so on. 1.2 Matching The intuitive definition of a matching is the specification of who marries whom : given a man of index m {1,..., N}, it is simply the index of the woman he marries, w = σ (m) {1,..., N}. Imposing that each man be married to one and only one woman at a given time translates into the requirement that σ be a permutation of {1,..., N}. This definition is too restrictive in so far as we would like to allow for some randomization. This could arise 9

13 because a given type is indifferent between several partner types; or because the analyst only observes a subset of relevant characteristics, and the unobserved heterogeneity induces apparent randomness. A feasible matching (or assignment) is therefore defined in all generality as a joint distribution Π over types of partners X and Ỹ, such that the marginal distribution of X is P and the marginal distribution of Ỹ is Q. ( ) We denote M P, Q the set of such joint distributions. 1.3 Surplus of a match The basic assumption of the model is that matching man m of full type x m and woman w of full type ỹ w generates a joint surplus Φ( x m, ỹ w ), where Φ is a deterministic function. Along with most of the matching literature, we assume that Assumption (O): Observability. Each agent observes the full types x and ỹ of all men and all women, but the econometrician only observes their components x and y. Assumption (O) rules out asymmetric information between participants in the market, as the economics of matching with incomplete information is a subject of its own. On the other hand, we do not really need to assume full information: Φ could for instance be reinterpreted as the expectation of a random variable conditional on x, ỹ, as long as all participants evaluate it in the same way. Matching markets are all about complementarities in the generation of surplus. If we are to identify such complementarities between observable types, we have to exclude complementarities between the unobserved components of full types. Following the insight of Choo and Siow (2006), formalized by Chiappori, Salanié, Tillman, and Weiss (2008), we therefore assume: 10

14 Assumption (S): Separability. Let x and x have the same observable type: x = x. Similarly, let ỹ and ỹ be such that y = y. Then Φ( x, ỹ) + Φ( x, ỹ ) = Φ( x, ỹ ) + Φ( x, ỹ). Assumption (S) requires that conditional on observable types, the surplus exhibit no complementarity across unobservable types. It is easy to see that imposing assumption (S) is equivalent to requiring that the idiosyncratic surplus from a match must be additively separable, in the following sense: Φ( x, ỹ) = Φ(x, y) + χ ( x, y) + ξ (ỹ, x) ( ) ) with E P (χ X, y X = x) 0 and E (Ỹ Q(ξ, x Y = y) 0 for every (x, y). Given Assumption (S), we call Φ(x, y) the observable surplus. Note that the model is invariant if one rescales the three terms on the right-hand side by the same positive constant. Later on we will normalize these three components. As proved in Chiappori, Salanié, Tillman, and Weiss (2008), assumption (S) implies that at the optimum (or equilibrium), a given individual (say, a man x) has a preference ξ ( x, y) for a particular class of observable characteristics (say y), but he is indifferent between all partners which have the same y but a different ỹ. More precisely, the optimal matching is characterized by two functions of observable characteristics U(x, y) and V (x, y) that sum up to Φ(x, y) such that if a man x is matched with a woman of characteristics ỹ, he will get utility while his match gets utility U(x, y) + χ( x, y) V (x, y) + ξ(ỹ, x). Chiappori, Salanié, Tillman, and Weiss (2008) showed that given assumption (S), the matching problem boils down to a set of single-agent choice problems for each type of man and of woman: for instance, man x is matched in equilibrium to a woman ỹ whose observable 11

15 type y maximizes U(x, y) + χ( x, y) over all values in the support of Q. One justification for assumption (S) would be that in a hypothetical match between man x and woman ỹ that results in a transfer of t from the man to the woman, the man gets utility U 0 (x, y) + χ 0 t and the woman gets V 0 (x, y) + ξ 0 + t with the restrictions that χ 0 + ξ 0 = χ( x, y) + ξ(ỹ, x) and U 0 + V 0 = Φ. Note that because transfers are endogenous at the optimum, U and V may be quite different from U 0 and V 0. It may be useful to resort to an analogy with the specification of demand systems, as used for instance in empirical industrial organization. The utility a consumer with observed type x and full type x gets from consuming a product with observed characteristics y and full characteristics ỹ can be decomposed into a sum of four terms: U 0 (x, y) + χ( x, y) + ξ(ỹ, x) + ζ( x, ỹ). The first one describes the average taste for observed product characteristics among consumers of a given observed type; the second one allows for unobserved variation in taste for observed characteristics; the third one allows for unobserved product effects; and the fourth one is the idiosyncratic term. Assumption (S) rules out this last term. Its strength depends on the quality of the data. There is clear evidence, for instance, that American couples produce more surplus when the partners have similar religious background. If religious affiliation is not in our dataset, then assumption (S) will not hold. Physical characteristics are a more subtle case. If the dataset contains no information on them, then assumption 12

16 (S) does not rule out a preference for good looks, nor variation in such preferences; but it does rule out a correlation between the preference for good looks and one s own good looks. While assumption (S) is already quite powerful, it still allows for very complex patterns: the covariance matrix of the χ( x, y) for a given man x is an unwieldy object not to mention other distributional characteristics. To go further, we need to add more restrictions on the specification of the components of the idiosyncratic surplus χ ( x, y) and ξ (ỹ, x). 1.4 Specifying the idiosyncratic surplus Following Choo and Siow (2006) and Chiappori, Salanié, Tillman, and Weiss (2008), we introduce the following assumption 2 : Assumption GUI: Gumbel Unobserved Interactions 1. The distributions of observed types P and Q are discrete, with probability mass functions p (x) and q (y) 2. There are an infinite number of individuals with a given observable type in the population 3. Fix the observable characteristics x of a man, and let ( y 1,..., y n Q) be the possible values of the observable characteristics of women. Then the preference shocks χ ( x, y 1),..., χ ( x, y n Q) are distributed as n Q independent and centered Gumbel (type- I extreme value) random variables with scale factor σ 1 ; similarly, 4. Fix the observable characteristics y of a woman, and let ( x 1,..., x n ) P be the possible values of the observable characteristics of men. Then the preference shocks ξ ( ỹ, x 1),..., ξ (ỹ, x n P ) are distributed as n P independent and centered Gumbel random variables with scale factor σ 2. 2 We define the scale factor to be 1 for the standard type-i extreme value distribution, which has variance π 2 /6; thus e.g. χ has variance σ1π 2 2 /6. 13

17 (GUI) underlies the standard multinomial logit model of discrete choice. We use it for the Independence of Irrelevant Alternatives property: without it, the odds ratio of the probability that a man with observable type x ends up in a match with a woman of observable type y rather than with z would also depend on the types of other women, and the model would become unmanageable. This assumption has well-known limitations. The first one is that it does not extend directly to continuous choice. We are currently exploring alternative specifications that would allow us to deal with continuous characteristics. It also restricts both heteroskedasticity and correlation patterns. We discuss extensions in section 8. Finally, part 2 of assumption (GUI) is made strictly for notational simplicity: it allows us to replace averages with expectations. If there are, say a finite number m of members of each observed type, then our main results in the next two sections only hold asymptotically in m. When we describe our estimators in section 6, we of course take into account the fact that we only have a finite sample. Under assumptions (O), (S), and (GUI), the model is fully parametrized; its parameters can be collected in a vector θ = (Φ, σ 1, σ 2 ), where Φ is the observable surplus function and σ 1 (resp. σ 2 ) is the scale factor of the unobservable characteristics of the men (resp. of women). Without loss of generality, all components of θ can be multiplied by any positive number; hence we shall need to impose some normalization on θ. We return to this later. As we will see, the total heterogeneity (σ 1 + σ 2 ) plays a key role in our results; thus we introduce a specific notation for it: σ = σ 1 + σ 2. 14

18 2 Solving for the Optimal Matching In this section we assume (O), (S), and (GUI), and we consider the problem of optimal matching: W(θ) = sup ( ) E Π Φ X, Ỹ. (2.1) Π M( P, Q) As the tilde signs in the formula suggest, none of the relevant quantities is observed; our main aim in this section is to prove that the formula can be rewritten entirely in terms of observable quantities, making inference possible. We examine first the dual, and then the primal version of the problem. 2.1 The Dual Let us provide some intuition before we state a formal theorem. Under (O), (S) and (GUI), standard formulæ for the multinomial logit model give the expected utility of a man of observable type x at the optimal matching: [ E max y ( U(x, y) + χ( X, ) ] y) X = x = σ 1 log y exp (U(x, y)/σ 1 ). Therefore the expected social surplus from the optimal matching is simply 3 (adding the equivalent formula for women of observable type y): σ 1 E P log y exp(u(x, y)/σ 1 ) + σ 2 E Q log x exp(v (x, Y )/σ 2 ). Now recall that U(x, y) is the mean utility of a man with observable type x who ends up being matched to a woman with observable type y at the optimum. As in the general development of the theory of matching, U is the value of the multiplier of the population constraints; and as such, it (along with V ) is the unknown function in the dual program 3 Since this formula may not be entirely transparent, we develop one term below: E P log X y exp(u(x, y)/σ 1) = X x p(x) log X y exp (U(x, y)/σ 1). 15

19 in which the expression for the social surplus above is minimized over all U, V such that U + V Φ. We now state this as a theorem (proved in Appendix B): Theorem 1 (Social welfare: dual version) Assume (O), (S), and (GUI). Then ( ) W(θ) = inf (U,V ) A σ 1 E P log y exp(u(x, y)/σ 1 ) + σ 2 E Q log x exp(v (x, Y )/σ 2 ) (2.2) where the constraint set A is defined by the inequalities x, y, U(x, y) + V (x, y) Φ (x, y). At an optimal matching, men with observable type x will be found in matches with women with observable types y such that U(x, y) + V (x, y) = Φ (x, y). The expected utility of men with observable type x matched with women of observable type y is U(x, y). 2.2 The Primal Theorem 1 also has a primal version, of course; and it is in fact our most useful result, as it will lead directly to an empirical strategy. While deriving the theorem takes a bit more work (again, see Appendix B), the intuition is simple. First, if there were no unobserved heterogeneity (with σ close to zero) the optimal matching would coincide with the optimal observable matching Π, which solves W(θ) = sup E Π Φ (X, Y ). Π M(P,Q) Going to the polar opposite, in the limit when σ goes to infinity only unobserved heterogeneity would count; and since it is just noise, the optimal matching would simply assign partners randomly, yielding the product measure P Q. As it turns out, when σ takes any intermediate value the optimal matching maximizes a weighted sum of these two extreme cases: 16

20 Theorem 2 (Social welfare: primal version) Under the assumptions of Theorem 1 ( ) W(θ) = π(x, y)φ (x, y) σi (Π) + σ 1 S(Q) + σ 2 S(P ), (2.3) sup Π M(P,Q) x,y where S (P ) and S (Q) are the entropies of P and Q given by S(P ) = x p(x) log p(x); and S(Q) = y p(y) log p(y); and I(Π) is the mutual information of the joint distribution Π, given by I(Π) = x,y π(x, y) log π(x, y) p(x)q(y). The mutual information I (Π) is just the Kullback-Leibler divergence of Π from the independent product P Q to Π. It is easy to see that I is a strictly convex function of Π. Moreover, I(Π) = S(P ) + S(Q) S(Π); and since Π has marginals P and Q, 0 S(Π) S(P ) + S(Q), so that we also have 0 I(Π) S(P ) + S(Q). The left hand-side is an equality at any pure matching (when all π s are 0 or 1), and the right hand-side inequality becomes an equality when where Π = P Q. Mutual information is a measure of the covariation of types x and y. Now P Q is the independent product of P and Q, which corresponds to a completely random matching Π = P Q. Thus a large positive I (Π) indicates that the matching Π induces strong correlation across types; I(Π) = S (P ) + S (Q) if and only if Π = P Q. If σ is very large then the Theorem suggests that I(Π) should be minimized, which can only occur for the independent matching Π = P Q; whereas if σ is negligible then Π should be chosen so as to maximize the expected observable surplus E Π Φ(X, Y ). This corroborates the intuition given earlier. 17

21 The optimal matchings coincide with the solutions to this maximization problem. Since we only observe the realized Π over observable variables, Theorem 2 defines the empirical content of the model: a combination of the parameters θ = (Φ, σ 1, σ 2 ) is identified if and only if the solution Π depends non-trivially on it. We already knew that θ can be rescaled by any positive constant without altering the solution. We can now go one step further: while all components of θ figure in this theorem, σ 1 and σ 2 only enter through their sum σ (the terms σ 1 S(Q) and σ 2 S(P ) do not depend on Π and therefore do not help for identification.) Thus and as announced, σ 1 and σ 2 are not separately identified: only the total heterogeneity σ is. Accordingly, we redefine the parameter vector θ as θ = (Φ, σ). Taking the limit when σ 0 in Theorems 1 and 2 and denoting W 0 (Φ) W(Φ, 0), we obtain as a corollary the classical duality of optimal matching: Corollary 1 (Homogeneous social welfare) Assume (O); then a) The value of the social optimum when θ = (Φ, 0) is given both by W 0 (Φ) = max Π M(P,Q) π(x, y)φ (x, y), (2.4) x,y and by W 0 (Φ) = inf (u,v) A 0 where the constraint set A 0 is given by ( p(x)u (x) + x y q(y)v (y) ) (2.5) x, y, u(x) + v(y) Φ (x, y) ; A matching (X, Y ) Π is optimal for Φ if and only if the equality u (X) + v (Y ) = Φ (X, Y ) holds Π-almost surely, where u and v solve the optimization problem (2.5). 18

22 Since all men with observable characteristics x have the same tastes in the homogeneous limit, they all obtain the same utility at the optimum. The utility U(x, y) becomes a function of x only, which we denoted u(x) above; and this is just the Lagrange multiplier on the population constraint π(x, y) = p(x) which is implicit in the notation Π M(P, Q). y 3 Nonparametric Identification The results in the previous sections give a very useful description of the optimal matchings, and they show that σ 1 and σ 2 cannot be identified separately. On the other hand, we have not provided a proof of identification of the remaining parameters yet. We now set out to do so. First note that if σ > 0 and since mutual information I is strictly convex, the objective function in Theorem 2 is strictly concave. Thus the optimal observable matching maximizes a strictly concave function over a compact convex set, and it must be unique 4. Now remember that given assumptions (O) and (S), there exist two functions U(x, y) + V (x, y) = Φ(x, y) such that the optimal matching obtains when man x maximizes U(x, y) + χ( x, y) over y and woman x maximizes V (x, y) + ξ(ỹ, x) over x. Now if π is the observable component of an optimal matching, and given assumption (GUI), U(x, y) = σ 1 log π(x, y) + u(x), where u(x) = σ 1 log y ( ) U(x, y) exp. σ 1 In the literature on discrete choice, u is called the inclusive value: here u(x) is the expected utility of a man of observed type x on the marriage market. Similarly, V (x, y) = σ 2 log π(x, y) + v(y). 4 Related results are given in Decker, Stephens, and McCann (2009). 19

23 Now U and V depend on θ and are not easy to characterize as we will see; but we know that they must sum up to Φ, so that Φ(x, y) = σ log π(x, y) + u(x) + v(y). In this formula u and v still depend on θ in a complex way; but they only appear in terms that depend only on characteristics of one partner. This implies that the surplus function Φ is identified up to an additive function of the form a (x) + b (y). To state this more formally, define the cross-difference operator as 2 F (x, y; x, y ) = ( F (x, y ) F (x, y) ) ( F (x, y ) F (x, y) ), for any function F of (x, y). Then: Theorem 3 (Cross-differences are identified up to scale) Assume (O), (S), and (GUI). Take any θ = (Φ, σ 1, σ 2 ) with σ = σ 1 + σ 2 > 0. Then 1. There exists a unique observable matching π which maximizes the social welfare (2.3). 2. There exists a unique 4-tuple (π, u, v, c) that solves the following system: π (x, y) p (x) q (y) exp π M (P, Q) ( ) Φ(x,y) u(x) v(y) c σ, E P u (X) = E Q v (Y ) = 0. (3.1) u(x) and v are both finite-valued functions, and the constant c coincides with the value of the social welfare c = W(θ). 3. The probability π defined in 2. coincides with the optimal observable matching defined in As a consequence, 2 Φ σ 2 log π; and this is a necessary and sufficient condition for θ to rationalize π. 20

24 5. At the optimal matching, each possible match has positive probability: 0 < π(x, y) < 1 for all x, y such that p(x)q(y) > 0. Given Theorem 3, the complementarity of various components of the observable types (x, y) of the partners can be tested directly on log π, since 2 log π and 2 Φ have the same sign. Moreover, the relative strengths of complementarities along several dimensions (say education and income on example (ER)) at a point (x, y) can be estimated by evaluating 2 log π for values of (x, y ) that differ from (x, y) along these dimensions. These results are reminiscent of those in Fox (2009), although we obtained them under quite a different set of assumptions: we do not use variation across subpopulations, neither does his rank-order condition apply to our model. Note also that when specialized to onedimensional types, our result yields that of Siow (2009), who tests complementarity of the surplus function by examining log-supermodularity of the match distribution. Theorem 3 immediately gives us an estimator of the observable joint surplus function Φ: ˆΦ(x, y) = log ˆπ(x, y), with ˆπ an estimator of π. This estimator corresponds to an assumed σ = 1; the last part of the Theorem shows that multiplying it by any positive factor (σ) and adding any pair of functions of x and of y would also yield a perfectly valid estimator. The positive scale factor σ is obviously irrelevant; the indeterminacy up to additive functions of x and y may seem more surprising. These additive functions represent the expected utilities of men of observed type x and women of type y on the marriage market, which could only be identified by relating the proportion of individuals to remain single to their types; since we are focusing on the population of matched individuals, we cannot identify them. While u(x) and v(y) can be chosen arbitrarily, U(x, y) and V (x, y) are partially identified. To be more precise, recall the equations U(x, y) = σ 1 log π(x, y) + u(x) and V (x, y) = σ 2 log π(x, y) + v(y); 21

25 the extent to which U and V are identified is directly implied by the fact that π is identified but neither σ 1, nor σ 2, nor u or v are. Therefore only ratios of the form U(x, y 1 ) U(x, y 2 ) U(x, y 3 ) U(x, y 4 ) are point identified, with the obvious analog statement for V. Note that this only makes sense if σ 1 > 0 and π(x, y 3 ) π(x, y 4 ); while the latter is directly testable, the former is not. These results have another surprising consequence: if Φ, U, V rationalize the data, then for any µ = (µ 1, µ 2 ) 0, the linear combination Φ µ (x, y) µ 1 U(x, y) + µ 2 V (x, y) and the functions U µ µ 1 U, V µ µ 2 V also rationalize the data. This is a by-product of assumptions (S) and (GUI). 4 Empirical and computational strategies Theorem 3 and its corollary immediately suggest a very simple nonparametric approach. In this discrete case, a nonparametric estimator ˆπ N (x, y) is readily obtained, by counting the proportion of matches between a man of characteristics x and a woman of characteristics y. We can pick arbitrary functions a(x) and b(y) and a number σ > 0 and define ˆΦ N (x, y) = σ log ˆπ N (x, y) + a(x) + b(y), without any reference to basis functions imposing σ = 1 on the way. A nonparametric approach will often be unsuitable for applied purposes, when the aim is to test for stylized facts about the matching patterns. We could, however, take this nonparametric estimator as the basis for a parametric approach. 22

26 4.1 Parametric approach Suppose for instance that the researcher specifies a parametric family of observable surplus functions (Φ(x, y; β)). Then he may choose ( ˆβ, ˆσ) to minimize a distance ˆπ N π β,σ, with π β,σ the optimal matching given by Theorem 2. Since the problem is invariant to rescaling, some normalization (e.g. imposing the value of σ) will be required, unless it is already imposed by the parametrization. An alternative approach would start from a parametric specification of the surplus function as above and choose (β, σ) to maximize the likelihood function N log π β.σ (x i, y i ) i=1 where each observation i is a couple. This amounts, of course, to maximizing ˆπ N (x, y) log π β,σ (x, y), x,y where the sum now runs over all (x, y) cells. One problem with these two-step approaches is that they require solving for the optimal matching for potentially large populations, and a large number of parameter vectors during optimization. This may seem to be a forbidding task: there exist well-known algorithms to find an optimal matching, and they are reasonably fast; but with large populations the required computer resources may still be large. Fortunately, it turns out that introducing (our type of) heterogeneity actually makes computing optimal matchings much simpler. 4.2 Computation Choose a parameter vector θ = (Φ, σ) and return to the characterization of optimal matchings in equation 2.3. Dividing by σ and taking the logarithm, optimal matchings can also be obtained by solving the following minimization program: π(x, y) log min Π M(P,Q) x,y π(x, y) p(x)q(y) exp(φ(x, y)/σ). 23

27 Now define a set of probabilities r by r(x, y) = p(x)q(y) exp(φ(x, y)/σ) p(x)q(y) exp(φ(x, y)/σ); x,y and note that given any choice of parameters θ and known marginals (p, q), the probability r itself is known. Determining the optimal matchings therefore boils down to finding the joint probabilities π with known marginals p and q which minimize the Kullback-Leibler distance to r: π(x, y) π(x, y) log r(x, y). (4.1) x,y Equivalently, we are looking for the Kullback-Leibler projection of r on M(P, Q). This is a well-known problem in various fields, and algorithms to solve it have been around for a long time. National accountants, for instance, use RAS algorithms to fill cells of a two-dimensional table whose margins are known; here the choice of r reflects prior notions of the correlations of the two dimensions of the table. These RAS algorithms belong to a family called Iterative Projection Fitting Procedures (IPFP). They are very fast, and are guaranteed to converge under weak conditions. We only describe the application of IPFP to our model here; we direct the reader to Rüschendorf (1995) for more information. The intuition of equation 4.1 is quite clear: the random matching, which is optimal when σ is very large, has π(x, y) = p(x)q(y). For smaller σ s the probability of a match between x and y must increase with the surplus it creates, Φ(x, y); and given our assumption (GUI) on the distribution of unobserved heterogeneity, it should not come as a surprise that the corresponding factor is multiplicative and exponential. To describe the algorithm, we split π into 5 π(x, y) = r(x, y) exp( (u(x) + v(y))/σ). The functions u and v of course will only be determined up to a common constant. The algorithm iterates over values (u k, v k ). We start from u 0 σ log p and v 0 0. Then at 5 It can be shown that at the optimum π(x, y) = 0 where r(x, y) = 0. 24

28 step (k + 1) we compute exp( v k+1 (y)/σ) = q(y) x r(x, y) exp( uk (x)/σ) and exp( u k+1 (x)/σ) = p(x) y r(x, y) exp( vk+1 (y)/σ). Two remarks are in order here: first, we could just as well start from u 0 0 and v 0 = σ log q and modify the iteration formulæ accordingly. Second and just as in other Gauss- Seidel algorithms, it is important to update one component based on the other updated component: the right-hand sides have u k and v k+1. If (u, v) is a fixed point of the algorithm, then π(x, y) p(x)q(y) = exp ( Φ(x, y) u(x) v(y) Comparing this formula to Theorem 3 shows that u(x) and v(y) have a simple interpretation: they represent (up to a common additive constant) the expected utilities of a man of observable characteristics x and of a woman of observable characteristics y. Thus the IPFP algorithm gives us not only the optimal matching, but also these expected utilities. Of course, their values depend on the normalization of θ = (Φ, σ), both because of the scale factor σ and because Φ could be translated by a sum of a function of x and a function of y without changing the optimal matching. The formulæ above can be simplified further. Given data on N couples, the marginal p assigns 1/N probability to each of (x 1,..., x N ), and similarly for women. Define a matrix Ψ by Ψ ij = exp(φ(x i, y j )/σ), and vectors a k i = exp( uk (x i )/σ), b k j = exp( vk (y j )/σ). Then we end up with the shockingly simple and inexpensive formulæ for the IPFP algorithm: b k+1 = N Ψ a k and ak+1 = N Ψb k+1. Once the iterations have converged to some (a, b), the optimal matching is simply σ π ij = 1 N 2 Ψ ija i b j. ). 25

29 As will be shown in our application in section 7, the IPFP algorithm is remarkably fast. Yet it cannot substitute for the limitations of the data: even with large datasets, nonparametric estimation must face the fact that there are many possible (x, y) cells. Take the education-income example, and assume that we distinguish three levels of education and five income classes. Then x and y can each take 15 different values, and there are 15 2 = 225 cells. For some of these cells, the estimator ˆπ N will be zero; but more importantly, it will likely be rather imprecise in general, and so will any parametric estimator obtained by the minimum distance method described above. Among all parametric specifications of the observable surplus function Φ, linear models are the most natural. As we will see in the next section, they also yield both illuminating insights into the properties of the optimal matchings and an alternative, very appealing estimation method. 5 The Semilinear Case In this section, we assume that the analyst has chosen K basis assorting functions φ 1 (x, y),..., φ K (x, y); and that she specifies the observable surplus function Φ Λ (x, y) as a linear combination of these basis assorting functions, with unknown assorting weights Λ R K : Model (SLOI): Semilinear Observable Interactions. observable surplus function as The analyst specifies the K Φ Λ (x, y) = Λ k φ k (x, y) (5.1) k=1 where the sign of each Λ k is unrestricted, and not all are zero. Note that in the discrete case which we focus on in this paper, this specification is only restrictive if K is small enough. Indeed, choosing K = n P n Q and the family of basis 26

30 functions φ ij (x, y) = 1 {x=xi,y=y j } for i = 1,..., n P and j = 1,..., n Q generates all possible surplus functions. In most applications, the analyst will want to choose a value of K that is much smaller than n P n Q, and so (SLOI) does restrict the specification. To return to the education/income example (ER): we could for instance assume that a match between man m and woman w creates a surplus that depends on the similarity of the partners in both education and income dimensions. The corresponding specification would be (with education levels E = (D, G) coded as (0, 1)): Φ(x m, y w ) = e m=0,1;e w=0,1 Λ em,e w 1(E m = e m, E w = e w ) + i=1,...,n r;j=1,...,n r Λ ij 1(R m = i, R w = j). This specification only has (n 2 r + 4) parameters, while an unrestricted specification would have 4n 2 r. Such an unrestricted specification would for instance allow the effect of matching partners in income class 3 to depend on both of their education levels. An even more restrictive, diagonal specification would be Φ(x m, y w ) = e=0,1 Λ E e 1(E m = E w = e) + i=1,...,n r Λ R i 1(R m = R w = i). In this last form, it is clear that the relative importance of the Λ s reflects the relative importance of the criteria. Thus Λ R i are both in income class i, while Λ E 0 measures the preference for matching partners who measures the preference for matching dropouts. The relative values of these numbers indicate how social preferences value complementarity of incomes of partners more, relative to complementarity in educations. In model (SLOI), the set of parameters becomes θ = (Λ, σ). If the model is correctly specified, then Theorem 3 gives us an estimator of the assorting weights Λ and the total heterogeneity σ 6. In fact, the cross-difference operator is linear and so under (SLOI), 2 log π = 2Φ σ = K k=1 Λ k σ 2φ k ; 6 Recall that σ 1 and σ 2 are not separately identified. 27